CN106940887A - A kind of satellite sequence image clouds of GF 4 and shadow detection method under cloud - Google Patents

Abstract

Aiming at the radiation preprocessing application of the Gaofen-4 satellite image, especially the cloud and cloud shadow detection application, the invention provides a cloud and cloud shadow detection method of the Gaofen-4 satellite sequence image. For the sequence images of the same geographic area, different linear functions are used to achieve relative registration, and they are sorted according to the average surface radiance, and the linear relative radiation normalization is used to reduce the radiation difference caused by different acquisition times. According to the number of sequence images, select Based on the S-G filter algorithm, or based on the statistical automatic threshold method to mark clouds and cloud shadows, and correct the shadow pixel detection results according to the distance between the detected shadow pixel and the nearest cloud pixel. The key steps involved in the present invention are realized by mature algorithms, which have high stability and applicability. For the production of cloud and shadow under cloud detection products in the preprocessing of Gaofen-4 satellite data and the improvement of product accuracy, it provides Key technical support.

Description

A method for detecting clouds and shadows under clouds in GF-4 satellite sequence images

technical field

The invention relates to a remote sensing image radiation processing technology, in particular to a cloud and cloud shadow detection technology for high-resolution remote sensing satellite sequence images.

Background technique

Radiation preprocessing of remote sensing images has always been one of the main topics in remote sensing data processing. After the new remote sensing satellites are put into use, the preprocessing of the new data is an urgent issue. The preprocessing of remote sensing images usually includes geometric preprocessing and radiometric preprocessing. In addition to the key radiometric calibration, the statistics of image cloud cover is also an important step in radiometric preprocessing. According to the needs of satellite data users, the proportion of cloud cover has become an important indicator for selecting satellite images. For example, images with as little cloud cover as possible are used in classification and mapping applications, and more attention is paid to cloud coverage in meteorological and disaster reduction applications. Lots of images. At present, most remote sensing satellite image data products contain cloud amount information, and many of them contain special cloud coverage marking bands, which are convenient for users to distinguish clouds and land surfaces pixel by pixel. There are many cloud detection methods for remote sensing satellite images. Developing suitable cloud detection algorithms according to the characteristics of satellite data is an important step in data preprocessing. Simple histogram statistics combined with an automatic threshold method, or an automatic threshold method based on cloudless surface reference data, such methods are prone to false detections for snowy and bright dry surfaces; observations with a high degree of quantification including infrared bands Satellite data mostly rely on the low-temperature characteristics of cloud layers to detect clouds, and generally have high accuracy, and data products often include special cloud marker bands.

The Gaofen-4 satellite (hereinafter referred to as GF-4) is a geosynchronous orbit satellite launched by China in December 2015. It is equipped with a panchromatic and multispectral camera with a spatial resolution of 50 meters and a medium-wave camera with a resolution of 400 meters. The infrared camera adopts area array staring method for imaging, and the imaging interval is as fast as 20 seconds, which has the advantages of high time and high spatial resolution. Since the release of the first batch of images by the National Defense Science, Technology and Industry Bureau on February 3, 2016, GF-4 has acquired a large amount of data in China and surrounding areas, playing an important role in detecting forest fires, floods and other disasters. The preprocessing of GF-4 satellite images also includes two parts: geometry and radiation. Geometric preprocessing includes system imaging model construction, control point matching and geometric fine correction, etc. The goal is to achieve pixel-by-pixel registration of imaging data under the same map projection . Radiation processing includes radiation calibration, detection of clouds and shadows under clouds, etc. The goal is to make image pixel values accurately describe the surface radiation conditions.

The research and development of GF-4 satellite data preprocessing technology is based on making full use of the previous satellite data processing results, and also considers the characteristics of GF-4 satellite image itself, and develops a special processing algorithm. According to the characteristics of GF-4 satellite data, there are two main technical ways to realize cloud detection: one is to use the low temperature characteristics of clouds, and detect according to the characteristics of clouds with low brightness in the mid-wave infrared band and high brightness in the visible light band. Many methods are used; the second is to use the motion characteristics of clouds to detect clouds from sequence images. Since the GF-4 satellite adopts a geosynchronous orbit and uses an area array staring imaging method, it is easy to obtain a large number of images in the same geographical area. Motion characteristics, the cloud and the shadow under the cloud can be distinguished by using the sequence image.

Through the analysis and experiments on the specific data of GF-4, it is found that the mid-wave infrared data is difficult to be used for fine cloud detection. Pixel blocks, a total of 64 visible light pixels with a resolution of 50 meters; the geographical coverage does not overlap and the pixel size is different, the mid-wave infrared is 1204×1024 pixels, and the visible light is 10240×10240 pixels; there is a difference in imaging time, mid-wave infrared and visible light The images are designed to be imaged at a fixed interval of 45 seconds, and the movement characteristics of the clouds make the position and shape of the clouds in the two data different. In terms of data quality, there are still obvious noise problems in the imaging of mid-wave infrared cameras.

Aiming at the problem of cloud detection in GF-4 satellite images, in the case of difficulties in the application of mid-wave infrared images, it is an effective technical way to study the use of sequence images to improve the accuracy of cloud detection. In the actual disaster monitoring application of GF-4, The multi-day timing continuous observation data of the disaster area usually has strong sequence characteristics, which provides favorable conditions for cloud detection based on multiple/sequence images. The engineering production of radiation preprocessing products of GF-4 satellite data requires a robust algorithm for detecting clouds and shadows under clouds in sequential images.

Contents of the invention

The purpose of the present invention is to provide a sequence image cloud and cloud shadow detection technology for the application of GF-4 static satellite image preprocessing, especially for the L1 of panchromatic and multi-spectral images with 50 meters spatial resolution of GF-4 satellite The production of cloud detection products in level data products provides an algorithm process for producing marked cloud and cloud shadow band products. This technology is based on the mature remote sensing image relative radiation correction algorithm and S-G (Savitzky-Golay) filter algorithm, and according to the radiation preprocessing requirements of GF-4 satellite images and the radiation characteristics of sequence images, the fast cloud and cloud shadow detection algorithm process is customized.

The basic idea of the present invention is: for the sequence image data of the same geographical area acquired by GF-4 geostationary satellites, in the absence of a system geometric model, the linear function obtained by automatic image matching is used to realize the pixel-by-pixel relative positional relationship between the images Registration, using automatic relative radiation correction to reduce the radiation difference caused by different imaging times between sequence images, using S-G filter combined with automatic threshold to correct the ground surface pixel by pixel in sequence images, and dividing the cloud and under-cloud by comparing the values before and after pixel correction Shadows, the final output is a single-band cloud and cloud shadow marker data for each image in the sequence.

The GF-4 satellite sequence image is limited to a panchromatic, multi-spectral image with a spatial resolution of 50 meters, and the difference in latitude and longitude of the four corner points of the image does not exceed ±0.3 degrees, which can be obtained under the staring mode of the GF-4 satellite The sequence of images may also be obtained on different days and arranged in sequence according to the acquisition time.

The GF-4 satellite sequence image cloud and the shadow detection method under the cloud provided by the technical solution of the present invention are characterized in that comprising the following implementation steps:

A data preprocessing, obtain sequence image relative registration linear parameters;

B linear relative radiation normalization, automatically extracting pseudo-invariant feature points between pairs of sequence images, and statistically comparing the radiation differences of pseudo-invariant feature points in each image to find data with large overall radiation differences for relative radiation correction ;

C sequence image cloud detection, according to the number of sequence images, select the algorithm based on S-G filter, or mark the cloud and the shadow under the cloud with the automatic threshold method based on statistics, and obtain the mask band data of the cloud and the shadow under the cloud for each image;

D Correct the detection result, first correct the shadow pixel according to the distance between the detected shadow pixel and the nearest cloud pixel, and then find out the data with problems in the detection result by sequentially superimposing and displaying the original image and the cloud detection result map to correct the cloud area, If there are multiple pieces of data with poor cloud detection results, these data are formed into a new sequence of images and the above cloud detection process is performed again.

The above-mentioned implementation steps are characterized in that:

The data preprocessing described in step A includes data integrity check, geographic coverage check, sorting by surface average radiance, and some preparatory initialization processes for program operation on the input sequence images. The specific process of obtaining the relative registration linear parameters of the sequence images is to use the approximate latitude and longitude coordinates of the four corner points of the images to determine the approximate relative positional relationship between the images for the sequence images of the same geographical area, and determine the automatic segmentation of the images according to the positioning error. The matching search range, through automatic matching, obtains the control point data and fits a linear function, and realizes relative registration between sequence images through different linear functions, without performing transformations such as interpolation and resampling on the images themselves.

The linear relative radiation normalization described in step B is used to reduce the radiation difference between two sequence images due to the difference in data acquisition time, and the relative radiation correction in the field of remote sensing is used between two adjacent images in the sequence technology, the IR-MAD (re-weighted Multivariate Alteration Detection transformation) change based on multi-source canonical correlation analysis is adopted here to automatically extract the pseudo-invariant feature points of the two images; the pseudo-invariant feature points of the images are statistically compared Point radiation difference, the comparison method used is to compare the mean value difference between an image in the sequence and the pseudo-invariant feature points extracted from the images before and after the sequence in the entire sequence, if the image is consistent with the sequence If there is a large difference in the mean value of the pseudo-invariant ground object points between the front and rear images, then relative radiation correction needs to be performed on the image; the relative radiation correction is performed on the data with a large overall radiation difference as described above, and the relative radiation correction adopts pseudo-invariant The linear function fitted by the feature points is completed.

The sequence image cloud detection described in step C adopts different detection methods according to the number of sequence images. When the number of images is greater than or equal to 10, an algorithm based on S-G filtering is selected, and the sequence images are firstly filtered pixel by pixel, and then according to each The comparison of pixel values before and after image filtering distinguishes clouds and shadows under clouds; when the number of images is less than 10, the mean and median values of the sequence images are directly counted pixel by pixel, and the difference between the pixel values of each image and the statistical mean and median values of regional clouds and cloud shadow; this step obtains the cloud and cloud shadow mask band data of each image;

Correct the detection result described in step D, first correct the shadow pixel according to the distance between the detected shadow pixel and the nearest cloud pixel, and according to the maximum distance between the shadow under the cloud and the cloud, for example, the GF-4 image can take a value of 500 or 1000 Prime number, for each detected shadow pixel, if no detected cloud pixel is found within the radius pixel range, it will be judged as a false detection, and the pixel will be removed from the shadow pixel; the original image will be displayed by sequential overlay Find out the problematic data in the detection results with the cloud detection result map and correct the cloud area. The problematic data described here is mainly for two situations. One is that there are many fragments and loopholes in the detection results. At this time, the computer morphology method is used The second is that there are a large proportion of obvious false detections in the detection results, such as misdetection of highlighted ground and water surfaces as clouds. At this time, it is necessary to re-detect the modified map. If there is any data, these data will be composed into a new sequence of images and re-processed for cloud detection.

Compared with the prior art, the present invention has the following characteristics: the present invention provides a solution for detecting clouds and shadows under clouds in GF-4 sequence images, and realizes fast registration of pixel-by-pixel positional relationships of GF-4 sequence images through linear functions , using linear relative radiation correction to reduce the radiation difference between the data, and using the S-G filtering results to divide the cloud and the shadow under the cloud. The algorithm has a high degree of automation, and the detection process of clouds and shadows under clouds does not require human-computer interaction. Users only need to check the final detection results and reprocess individual data. The key steps involved are realized by mature algorithms, which have high stability and applicability. It provides key technical support for the production of cloud and shadow under cloud detection products and the improvement of product accuracy in the preprocessing of GF-4 satellite data.

Description of drawings:

Figure 1 is a flow chart of cloud detection in GF-4 satellite sequence images

Figure 2 is a schematic diagram of S-G filtering and threshold cloud detection at a single pixel

Figure 3 is a schematic diagram of the morphological correction of the detection results

detailed description:

The idea of this technology is to use the motion characteristics of clouds and shadows under clouds in sequence images to realize the detection of clouds and shadows under clouds. The necessary conditions are: the data acquired by GF-4 satellite can easily form a sequence, and the same There is only a positional relationship between rotation and translation among the multiple images of the region. This necessary condition is satisfied for the GF-4 satellite image, because the GF-4 satellite adopts a geostationary satellite orbit, its position relative to the earth is fixed, and its imaging geometry is unchanged, including any image within the observable range of the earth. The geometric relationship from one point to the imaging point of the satellite sensor is fixed. The fixity of the position of the geostationary satellite ensures multiple observation images of the same area. When the coordinates of the image center point and the four corner points are the same, the system imaging geometric models of all images are the same, so that the system distortion of the image is the same as The spatial resolution is the same. In addition, the GF-4 satellite uses the area array staring method for imaging, and the imaging time is instantaneous. The imaging process of the image will not introduce new geometric distortion, and the shaking of the satellite and the sensor will hardly affect the imaging. Finally, GF-4 satellite imaging positioning accuracy is high. Through pointing control, GF-4 can realize free observation of China and surrounding areas, and can also use staring mode to continuously observe a fixed area with a width of 400 kilometers. Its positioning accuracy reaches ±0.1 degrees, making GF-4 satellites have The ability to acquire sequential images of the same fixed area. Since the release of the first batch of images by the National Defense Science, Technology and Industry Bureau on February 3, 2016, the GF-4 satellite has acquired a large amount of data in China and its surrounding areas, including a large amount of data that can form a sequence, including the GF-4 satellite staring mode. A sequence of images obtained at similar times on the same day, and images of the same area acquired on different days constitute a sequence.

The process of detecting clouds and shadows under clouds in GF-4 satellite sequence images by using the present invention is shown in Fig. 1 , which is now described in conjunction with the accompanying drawings.

Processing unit 111 data preprocessing, data preprocessing is aimed at GF-4 satellite sequence images, for visible light and near-infrared images with a spatial resolution of 50 meters released by GF-4 data, there are two ways to obtain sequence images: one is GF-4 -4 A sequence composed of multiple consecutive images acquired by the camera in the staring working mode on the same day and in a similar time; the second is multiple images of the same area acquired on different days in the non-staring mode. In the actual operation of the GF-4 satellite, it is easy to obtain the sequence image data composed of the second non-staring mode. Because disaster monitoring is an important task of GF-4, when a disaster occurs in a certain area, GF-4 can observe the fixed area for multiple days and multiple times within the range of latitude and longitude error of 0.1 degrees, thus forming a sequence of images. The data preprocessing algorithm program performs data integrity checks, geographic coverage checks, sequence sorting, and some preparation initialization processing on the input sequence images. The sequence sorting is not according to the time of data acquisition, but according to the average radiance of the surface of each image.

Sorting according to the average radiance of the ground surface in each image In the present invention, the subsequent processing of GF-4 geostationary satellite data is critical. Unlike sun-synchronous orbit satellites, which acquire images of the same geographical location of the earth at similar times, geosynchronous orbit satellites remain unchanged relative to the earth, and can choose to image at any time of the day. Imaging at different times during the day depends on the height of the sun. The overall image Radiation brightness is also different, such as the difference in image radiation between 8 o'clock in the morning and 12 o'clock in the noon. The specific sorting method is to count the histogram of the blue light band in the sequence image, and after excluding the values of the overly bright and overdark pixels, the mean value of the remaining pixels is used as the basis for sorting. The purpose of excluding overly bright and overly dark pixel values here is to filter out possible clouds and shadows under clouds, so as to ensure that the obtained average value represents the radiance of the surface as much as possible.

After the sequence images are sorted, since the present invention belongs to GF-4 data radiation preprocessing, the images used are the original data without systematic geometric correction, and the pixel-by-pixel registration of ground objects cannot be realized only by direct superimposition based on image positioning information. Here, the linear relative registration method is used. The specific process is to use the approximate latitude and longitude coordinates of the four corners of the image to determine the approximate relative positional relationship between the images for the sequence images of the same geographical area, and determine the automatic matching retrieval of image blocks according to the positioning error. Range, through automatic matching, the control point data is obtained and a linear function y=ax+b is fitted, and the relative registration of the sequence images is achieved through different linear function parameters a and b, without interpolation and resampling of the images themselves etc. transform.

If the preprocessed sequence image data is different day data acquired in the non-staring mode, the processing unit 112 needs to perform linear relative radiation normalization processing. This processing is a key step in the algorithm flow, and it is used to reduce the radiation difference between two sequence images due to the difference in data acquisition time. The relative radiation correction technology in the field of remote sensing is used between two adjacent images in the sequence. For An image in the sequence, compared with the mean difference between the pseudo-invariant feature points extracted from the images before and after the sequence in the entire sequence, if the image is invariant to the pseudo-invariant If the mean values of the object points are very different, the image needs to be corrected for relative radiation. Here, the IR-MAD change is used to automatically extract the pseudo-invariant feature points of the two images.

The IR-MAD transform is derived from the MAD transform proposed by Nielsen et al. (1998). In order to cover the changing pixels in the two-temporal image, the algorithm first forms a linear combination of pixel values in N channels of the two images. Random vectors X and Y are used to represent the pixel values screened out in the overlapping area of the target image and the reference image, respectively. According to the following transformation formula:

U＝a ^T X＝a ₁ X ₁ +a ₂ X ₂ +Λ+a _N X _N

V=b ^T Y=b ₁ Y ₁ +b ₂ Y ₂ +Λ+b _N Y _N

Among them, a _i and b _i are MAD coefficients, and the MAD transformation minimizes the positive correlation between U and V. Under the premise of obeying the constraint: Var(U)=Var(V)=1, define the MAD variable:

MAD=Var(U-V)=Var(U)+Var(V)-2cov(U,V)=2(1-corr(U,V))→Maximum

Minimizing the positive correlation coefficient corr(U,V) is a standard statistical procedure, the so-called generalized eigenvalue problem. Each component of the obtained MAD variable is orthogonal to each other and is an invariant of linear transformation. The reason why the present invention chooses the MAD transformation to extract the invariant feature points is that the MAD transformation is not sensitive to the linear relationship between the variables X and Y, and can be well adapted to the GF-4 images acquired at different times There is a large radiation difference between them. The IR-MAD transformation further improves the accuracy and stability of the MAD algorithm.

Using the pseudo-invariant feature points of the two images automatically extracted from the IR-MAD changes, an overall linear function y=ax+b is fitted by the least squares method, and the sequence is transformed using the traditional linear relative radiation correction of remote sensing images An image with a large difference in radiance is corrected to the radiance level of an adjacent image.

According to whether the sequence of images contains more than 10 images, it is determined that the processing unit 113 or the processing unit 114 is used for subsequent processing.

Processing unit 113 statistics and automatic threshold cloud detection, the specific algorithm is, if the sequence image contains n images, n≤10, for each relative registered pixel position in the sequence image, count the mean value Vmin and median value of n pixels Vmid, if there is little difference between the mean value and the median value, such as |Vmin-Vmid|<10, then all n pixels are marked as the surface. If the difference between the median value and the mean value is large, compare n pixels with Vmid one by one. If Vi-Vmid>Vcloud, then determine that the i-th pixel is a cloud, and Vcloud is the cloud threshold; if Vi-Vmid<Vshadow, then determine The i-th pixel is the shadow under the cloud, and Vshadow is the shadow threshold under the cloud, which is a negative value; the possible values of Vcloud and Vshadow are ±2|Vmin-Vmid|, or ±3|Vmin-Vmid|.

Processing unit 114 S-G filtering and threshold cloud detection. This step is to first perform S-G filtering on the sequence image, and then determine whether it is a cloud or a shadow under the cloud by comparing the difference between the values before and after filtering.

S-G filter achieves the purpose of smoothing sequence data through polynomial fitting of sliding window (Savitzky & Golay, 1964). The number of sequences is N, and the k (k≤n) order polynomial fitting for the subsequence of which the length is n=2m+1 can be expressed as:

The SG filtering process is to perform k-order (k≤n) on a certain point t ₀ in the sequence and a total of n=2m+1 points (ti, yi) in its left and right m neighborhoods, i∈[-m, m] Polynomial fitting, replace the data (t ₀ , y ₀ ) in the original time series with the data (t ₀ , y ₀ ) in the center of the fitted sliding window, and then move the window to the right so that the center of the window moves to the bottom of the sequence 1 data, repeat the above process until the sliding window reaches the end of the sequence. The smoothing window coefficient is obtained by the method of least squares.

If the sequence image contains n images, n>10, for each relatively registered pixel position in the sequence image, after SG filtering, compare n pixels V _i with the filtered value V _i-SG successively, If V _i -V _i-SG ＞V _cloud , it is determined that the i-th pixel is a cloud; if V _i -V _i-SG <V _shadow , then it is determined that the i-th pixel is a shadow under the cloud; V _cloud and V _shadow The empirical value is desirable, such as 20 or 30, and the threshold can be adjusted according to the specific data situation. The schematic diagram of SG filtering and threshold cloud detection at a single pixel in a sequence of images is shown in Figure 2.

Correct the detection result of the shadow under the cloud, correct according to the distance between the detected shadow pixel and the nearest cloud pixel, and give the maximum distance between the shadow under the cloud and the cloud. For example, the GF-4 image can take a prime number of 500 or 1000, For each detected shadow pixel, if no detected cloud pixel is found within the radius pixel range, it is judged as a false detection, and the pixel is removed from the shadow pixel.

The accuracy of the detection results is recommended to be checked manually, and the results with poor accuracy can be processed again after adjusting the threshold according to the specific data situation. In addition, because the cloud detection is based on a single pixel, sometimes there may be a large number of fragments and holes in the local area of the detection result. In order to improve the boundary effect of cloud detection, the method of computer morphology is used for cloud area trimming processing, removing the isolated cloud area less than a certain number of pixels outside the cloud boundary, filling the hole less than a certain number of pixels in the cloud boundary, and then trimming the cloud boundary . The schematic diagram of the effect is shown in Figure 3. Since the actual cloud itself may also be overly fragmented, it is up to the user whether to groom the cloud zone.

The detection results of clouds and shadows under clouds are saved as 8-bit single-band images, with the value of 0 for the surface, 1 for shadows under clouds, and 2 for clouds. N pieces of sequence images correspond to n pieces of detection results, which are provided to users as GF-4 primary data products.

The example of the present invention has been implemented on the PC platform, and has been delivered to the user for testing and use at present, as a key technology for inversion of cloud characteristic parameters in GF-4 data radiation preprocessing.

It should be pointed out that the specific embodiments described above can enable those skilled in the art to understand the present invention more comprehensively, but do not limit the present invention in any way. Therefore, those skilled in the art should understand that the present invention can still be modified or equivalently replaced; and all technical solutions and improvements that do not depart from the spirit and technical essence of the present invention should be covered by the protection scope of the patent of the present invention.

Claims (4)

1. a GF-4 satellite sequence image cloud and shadow detection method under the cloud, the method is aimed at Gaofen No. 4 satellite image radiation preprocessing application, particularly cloud and shadow detection application under the cloud, it is characterized in that comprising the following implementation steps: A data preprocessing, obtain sequence image relative registration linear parameters; said data preprocessing is sorted according to surface average radiance; obtain sequence image relative registration linear parameters, the specific process is to sequence images of the same geographical area, using image four The approximate latitude and longitude coordinates of each corner point determine the approximate relative positional relationship between the images, and determine the retrieval range of the automatic matching of image blocks according to the positioning error. Through automatic matching, the control point data is obtained and a linear function is fitted. Realize relative registration through different linear functions, without performing transformations such as interpolation and resampling on the image itself; B linear relative radiation normalization, automatically extracting pseudo-invariant feature points between pairs of sequence images, and statistically comparing the radiation differences of pseudo-invariant feature points in each image to find data with large overall radiation differences for relative radiation correction ; The linear relative radiation normalization reduces the radiation difference caused by the difference in data acquisition time between the sequence images, and the relative radiation correction technology in the field of remote sensing is used between adjacent two images in the sequence; The radiation difference of image pseudo-invariant feature points, the comparison method used is the size of the mean difference between an image in the sequence and the pseudo-invariant feature points extracted from the images before and after the sequence in the entire sequence In comparison, if the average value of the pseudo-invariant surface object points between the image and the previous and subsequent images in the sequence is very different, then the image needs to be corrected for relative radiation. C sequence image cloud detection, according to the number of sequence images, select the algorithm based on S-G filter, or the automatic threshold method based on statistics to mark clouds and cloud shadows, and obtain the cloud and cloud shadow mask band data of each image. 2. The method according to claim 1, characterized in that: For sequence image cloud detection, different detection methods are adopted according to the number of sequence images. When the number of images is greater than or equal to 10, the algorithm based on S-G filtering is selected, and the sequence images are first filtered pixel by pixel, and then the pixel values before and after filtering are used for each image. The comparison distinguishes the cloud and the shadow under the cloud; when the number of images is less than 10, the average and median values of the sequence images are directly counted pixel by pixel, and the cloud and the shadow under the cloud are based on the difference between the pixel value of each image and the statistical mean and median. 3. The method according to claim 1, characterized in that: Statistical and automatic threshold cloud detection, the specific algorithm is, if the sequence image contains n images, n≤10, for each relative registered pixel position in the sequence image, count the mean value Vmin and median value Vmid of n pixels, if There is not much difference between the mean value and the median value. For example, when |Vmin-Vmid|<10, all n pixels are marked as the surface; -Vmid>Vcloud, then determine that the i-th pixel is a cloud, and Vcloud is a cloud threshold; if Vi-Vmid<Vshadow, then determine that the i-th pixel is a cloud shadow, and Vshadow is a cloud shadow threshold, which is a negative value; Vcloud and The possible value of Vshadow is ±2|Vmin-Vmid|, or ±3|Vmin-Vmid|. 4. The method according to claim 1, characterized in that: Correct the detection result, correct the shadow pixel according to the distance between the detected shadow pixel and the nearest cloud pixel, and according to the maximum distance between the shadow under the cloud and the cloud, for example, the GF-4 image can take a value of 500 or 1000 prime numbers, for each detection If there is no detected cloud pixel within the radius pixel range, it will be judged as a false detection, and the pixel will be removed from the shadow pixel.